v0.8.0

Fixed

• Typed entries: parse_entries() now returns ErrorEntry dataclasses (dict-compatible via __getitem__), full type hints on all functions, and a small exception vocabulary (AgentLogError / ValidationError / LockTimeoutError) — same behavior, same exit codes.

• stdin reconfigured to UTF-8 on Windows: piped unicode no longer double-encodes into the log (stdout-only reconfigure bug).

• L10: load() no longer crashes on a locked/unreadable log file (graceful OSError fallback; regression tests added).

• Concurrent --add appends no longer lose entries: the append is
  now serialized by a cross-process lock file (<log>.lock, stdlib-only,
  atomic O_CREAT|O_EXCL create with 5s wait and stale-lock recovery)
  and the log is re-read inside the lock before writing (lost-update fix).

• _extract_area() marker semantics documented and pinned by tests:
  the CI gate and the shell hooks agree on the marker (first matching
  line, last AREA:/LOG: on it — the hooks' grep -m1 + greedy sed).

• status_token() strips the en-dash as well as the em-dash (OPEN– ->
  OPEN).

• start.py active_errors() delegates to check_errors.parse_entries()
  (the canonical parser) instead of re-implementing entry splitting, so
  the boot briefing can never drift from the tool's format.

• load() reads with utf-8-sig so a BOM-prefixed log is parsed, not
  silently ignored.

Added

• Robustness tests: 100-entry fuzz, BOM / invalid UTF-8, en-dash
  statuses, an empty section-5-only log, and multi-marker precedence in
  --check-commit (last marker on the first matching line).

Docs

• Add a permanent visitor badge to the README (GitHub traffic stats
  only keep 14 days).
• README: document the lessons-clustering chain limitation, the parser's
  worst-case O(n²) on pathological logs, and LOGNAME repo-root-relative
  paths.